Conference Talk – OLA 2025 (Video)
Date:
At the International Conference on Optimization and Learning (OLA 2025), I presented our paper on addressing the Hyperparameter Optimization (HPO) challenge in Large Language Model (LLM) fine-tuning.
Our work investigates two complementary approaches:
- Bayesian Optimization with Gaussian Processes (BO-GP), which is efficient in leveraging prior knowledge but limited by its sequential nature.
- Partition-Based Optimization (PBO), which enables massive parallelization but requires more evaluations to converge.
By combining their strengths, we proposed a hybrid BO-PBO algorithm, paving the way for scalable optimization methods tailored to expensive objective functions in exascale computing environments.
📺 You can watch the presentation here or look at the slides here